Pixelwise Fat-Water Separation by Physics-guided Learning with Simulation Data

Motivation: Physics-guided learning has shown excellent performance in MRI field without the need for training data acquisition. The well-established mathematical model for fat-water separation allows the application of physics-guided learning. Goal(s): Propose a pixel-by-pixel approach for fat-water separation with deep neural network trained with simulation data only. Approach: Fat-water separation is performed using an artificial neural network trained on simulation data, with MR physics-based constraint. The network parameters are optimized to the target data for higher performance. Results: The proposed network trained only on simulation data is applicable to in vivo data. The physics-guided learning and optimization provides stable training and improved image quality. Impact: The proposed method allows to perform fat-water separation without the acquisition of training data yet successfully performs fat-water separation. Moreover, the pixelwise approach ensures fast training and inference with low computational cost.